Broadcom Forecasts 88-Fold Surge in Custom AI Chip Revenue by 2028
Broadcom projects its custom AI accelerator and networking chip revenue will leap from $2.6 billion in fiscal 2022 to $230 billion by fiscal 2028, a two-order-of-magnitude increase. The company has secured HBM memory, wafer supply, and packaging capacity to support this growth, with CEO Hock Tan telling analysts not to expect quarterly forecast updates. Key customers include Google, Anthropic, OpenAI, and Meta, while Amazon uses Broadcom's packaging for its Trainium 4 chip.
Broadcom's forecast spans six custom accelerator customers, with Google receiving "multi-tens of billions" in TPU shipments annually. Anthropic's deployment roadmap shows 1 gigawatt of Ironwood TPU 7 this year, expanding to 5 gigawatts of TPU 8i in 2027 and 10 gigawatts of TPU 9 by 2028, positioning Anthropic as Broadcom's largest XPU customer by 2027.
The company's cost advantage is notable: custom XPU systems run roughly $12-15 billion per gigawatt, compared to $18 billion for Nvidia's Grace-Hopper and $40 billion for Vera-Rubin platforms. Broadcom has locked in HBM memory, wafer, and packaging capacity to support the projected growth, with CEO Hock Tan declining to provide quarterly forecast revisions.
This forecast could reshape the AI infrastructure market by giving hyperscalers and AI labs a credible alternative to Nvidia's dominant GPU platforms. If Broadcom's projections hold, custom silicon may lower the cost of large-scale AI compute, potentially accelerating deployment of frontier models while concentrating market power among a handful of well-funded customers who can commit to gigawatt-scale deployments. Smaller players may find themselves priced out of the custom chip market.